The Truth Engine We Have Yet to Build

This is the author’s accepted version of an essay appearing in the Aftershock column of IEEE Computer, targeted for September 2026, DOI 10.1109/MC.2026.3707921. Posted per IEEE’s author self-archiving policy. Please cite the published version.

^{© 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works.}

Because today’s AI optimizes for human approval, it treats truthfulness as a tunable behavior rather than a constraint. This essay argues that truth must be evaluated upstream of alignment, via an epistemological world model beyond physics.

In 1666, Leibniz asked whether human reasoning could be mechanized—whether controversies could be resolved not by debate but through calculation. He later called it the calculus ratiocinator and dedicated the rest of his life to it, writing in his final year that he hoped to achieve “in all fields of inquiry capable of certainty, what algebra does in mathematics.”¹ In 1847, Boole provided that algebra, formalizing the essence of logical thinking itself.² In 1938, Shannon demonstrated that Boolean algebra mapped perfectly onto switching circuits—an insight that would eventually influence every digital circuit ever built.³

With artificial intelligence (AI), we have finally built Leibniz’s machine. It has not settled a single controversy.

It is tempting to dismiss this as a temporary symptom of a developing technology. After all, the modern generative AI (GenAI) systems that simulate reasoning have only recently entered the mainstream market. It’s easy to assume that it’s just a matter of time—as large language models grow, as context windows expand, and as hallucination rates decrease, these thinking machines will naturally enable the consensus Leibniz envisioned.

But evidence points in the opposite direction. The behaviors of current AI, particularly its widespread tendency toward sycophancy—where models learn to flatter users and confirm flawed premises to boost approval ratings—suggest that time and scale alone will not fix the problem. The reason our AI systems fail to resolve our disputes is not a lack of technological maturity; it’s a matter of architectural design. We built Leibniz’s machine, but we tuned it for a post-truth world.

Post-Truth: Emancipation Without Reconstruction

A decade ago, Oxford Dictionaries chose “post-truth” as its word of the year, acknowledging a cultural shift that had become impossible to miss: objective knowledge was losing ground to appeals to emotion and personal belief. That moment did not come out of nowhere. It marked the political and cultural triumph of a critique long cultivated by postmodernist philosophers. From Nietzsche⁴ to Foucault,⁵ these thinkers argued that truth claims, including scientific ones, are never innocent. They inherently encode and preserve political and economic power.

For a long time, this critique stayed safely within universities and elite intellectual circles, where it played a necessary and constructive role. It freed scholars from rigid, often oppressive historical narratives and prompted serious questioning of institutional biases. However, even within academia, this liberation occurred without rebuilding the foundations of reliable knowledge. Postmodernists excelled at dismantling inherited frameworks of consensus, but they provided no solid epistemological structure to replace them. As a result, they left a philosophical gap where the idea of “truth” once resided.

Then came the Internet, and that void spilled into the public sphere. To grasp how much is missing, we need to look at the last time a technology disrupted our consensus. When the printing press appeared in the 15th century, it caused a major epistemic collapse. By unleashing a flood of pamphlets that circumvented the traditional elite gatekeepers of publishing, such as the Church, it opened the gates for centuries of often violent upheavals. But society eventually experienced an epistemic rebuild. We created entirely new institutions to check and filter what was then an unprecedented amount of information—scientific journals, encyclopedias, editorial standards, journalistic ethics, and eventually, peer review. These new structures restored a shared understanding of truths while preserving the democratizing, empowering potential of the printing press.

The digital revolution, however, has produced no such reconstruction. As the postmodern dismantling of authority reached the public at large, the web and social media simply weaponized the void. Instead of building new institutions to evaluate digital knowledge, we replaced a shared, regulated body of knowledge with fragmented digital narrowcasting and algorithmic echo chambers.

Now, GenAI is operating in the same void. Because we never built a new, rigorous framework for evaluating truth after the postmodern critique dismantled the old one, our most advanced AI systems are defaulting to the only metric left: human preference.

Truth Is Upstream of AI Alignment

Over the past few years, the new crop of GenAI systems and tools has added a completely new twist to the epistemic crisis. Because chatbots and agents produce the most digested and least contextualized forms of knowledge, they sound incredibly authoritative. This raises the question of how these systems ensure the accuracy of their information.

Currently, the AI industry’s answer is “alignment.” The overarching focus of frontier labs is aligning models with human values, safety protocols, and user preferences. Within this paradigm, truthfulness is treated merely as a sub-component of alignment—a behavioral dial to be tuned alongside politeness, safety, and helpfulness.

We can see this hierarchical choice clearly in how leading labs operationalize “truth” today:

Anthropic (Claude): While Anthropic’s AI Constitution⁶ explicitly lists “being honest” as a core goal, it provides “heuristics for weighing helpfulness against other values.” Truth is placed on the same tuning scale as being thoughtful and caring, making it a coordinate value to be traded off rather than a foundational constraint.

Google (Gemini): Google treats truthfulness as a quantity to be measured, not a constraint to be enforced. Factuality is just one benchmark among many in Gemini’s model cards,⁷ balanced against tone, instruction-following, and “unjustified refusals.” Sycophancy is tuned as a “persona attribute” rather than structurally guaranteed. Formal, threshold-based governance is reserved for catastrophic risks.⁸ Truthfulness stays a line item on a model card, not a gate in the safety framework.

OpenAI (ChatGPT): OpenAI’s Model Spec⁹ explicitly emphasizes behavioral instructions and stylistic norms, mentioning that directives like “Be honest” must be balanced against directives like “Be warm,” and tends to prioritize helpfulness and compliance over strict factual accuracy.

xAI (Grok): Despite claiming that its agents are “maximally truth-seeking,”¹⁰ xAI never defines this philosophically, architecturally, or operationally. Without a rigorous framework, the claim acts as a marketing banner rather than a verifiable system obligation.

Subsuming truth as a component of alignment is an architectural weakness that degrades the models. The industry-wide reliance on reinforcement learning from human feedback (RLHF) provides the mechanism. Because these systems are optimized to maximize human approval, they routinely fall into the “sycophancy trap.”¹¹ They learn to flatter the user, affirm flawed premises, and agree with false assertions rather than risk pushback.

The core issue is the fundamental veracity of what the system generates. While labs are currently attempting to fine-tune sycophancy away by explicitly penalizing blind agreement, these efforts face an intrinsic limit. RLHF—and its AI-driven variant reinforcement learning from AI feedback (RLAIF)—maps outputs to a scalar reward based on human preference or predefined behavioral rubrics. This means the model is being optimized for persuasiveness and social acceptability, not factual accuracy. If an AI corrects a user on a trivial error, the reward mechanism might function well. But if the AI corrects a user on a deeply held, emotionally charged false belief, the system is structurally forced to weigh that factual accuracy against competing directives to be “warm,” “helpful,” or “safe.” A system engineered to mathematically balance the veracity of its statements against user placation will eventually compromise objective reality to maximize its reward.

These well-documented failures support my core assertion: truth must be upstream of alignment.

The very primacy of alignment in the current discourse reflects a post-truth bias, assuming that human preferences should take precedence over objective reality. In all current models, AI is ultimately aligned with the user rather than a stable body of knowledge. The resulting loss of common ground threatens our ability to cooperate effectively, let alone produce good results. The goal, then, is simply to help build a reliable common ground for successful coordination in various contexts.

The Necessary Inversion: Toward an Epistemological World Model

We need an architecture where truth is evaluated systematically first and alignment built safely on top of that evaluation.

In Sapiens, Yuval Noah Harari argues that human culture is uniquely characterized by our ability to create and believe in shared narratives; we shape the meaning of the world through the stories we share.¹² However, as David Deutsch shows in The Beginning of Infinity, the reliability of our cumulative evolution depends on “tight explanations”—theories that are difficult to vary and rigorously tested against reality.¹³

Today, we face an epistemic emergency because we have decoupled these two forces. Our technology has scaled our capacity to fabricate narratives—initially through social media and now through AI—by orders of magnitude. Meanwhile, our analytical capacity to verify the “tightness” of those narratives remains limited to manual, human-speed thinking.

A framework that prioritizes truth over alignment is necessary because current AI architectures seem incapable of bridging this gap. We cannot “prompt” our way out of the sycophancy trap, nor can we rely on rewards to foster human convergence on shared truths. As long as we treat truth as a behavioral dial within an alignment framework, we are automating the generation of persuasive but potentially incoherent narratives. We must have the option to force the AI to operate within the constraints of reliable knowledge, transitioning the system from a pure “narrative-generation engine” into one enhanced with a validation engine. This validation engine is best understood as an epistemological world model, generalized beyond physics to legal, moral, political, cultural, and even artistic realities.

To build this validation engine, we must look beyond the multibillion-dollar race to build world-model AI. A world model equips an AI with a simulator of its environment to predict the consequences of an action before making a move. The concept has deep research roots,¹⁴ originating with early neural networks designed to simulate their physical surroundings.¹⁵ Examples of this approach include Yann LeCun’s joint embedding predictive architecture (JEPA) framework¹⁶ and his startup, AMI Labs, as well as Fei-Fei Li’s World Labs startup.

While this new wave of world models correctly identifies the need to ground AI generation, the current scope of what “world” means remains far too narrow. The world model paradigm anchors AI primarily in a sensory-driven or mathematically tractable baseline, focusing on domains such as intuitive physics or spatial reasoning. But there are many other “realities”: legal frameworks, moral norms, psychological insights, and cultural narratives, to name just a few. We need a framework that is broad enough to cover all of these, albeit perhaps not as powerful as a full simulation of the world in question.

---

The postmodernists were right to dismantle the old, centralized monopolies on knowledge, successfully showing that “Truth” with a capital T often served as a mask for power. But their liberation left an opening, which the Internet and GenAI have since exploited. As we face this crisis, we must understand that post-truth should not, and cannot, mean no truth. If we give up the pursuit of reliable, shared knowledge, we hand over our digital public square entirely to raw power, algorithmic echo chambers, and the sycophancy trap.

To survive this era, we need an epistemic reset similar to what happened after the printing press. We urgently require effective scalable institutions to manage reliable shared knowledge. However, to honor the valid points of the postmodern critique, these institutions cannot give a small elite too much power, nor should they function as totalitarian gatekeepers. They must be mathematically rigorous, transparent in their structure, and capable of operating at the unprecedented scale and speed of digital life.

This is a formidable engineering challenge. But it is a challenge we must accept. GenAI may accelerate our current epistemic fracturing. But if we choose to engineer these systems to prioritize truth over mere behavioral alignment, they could become the tools that help us rebuild a shared world. And Leibniz’s machine might finally let us calculate our way to common ground.

Acknowledgment

I’m grateful to Bharat Shyam, Lila Shroff, Mourad Heddaya, and Shehab Heddaya for their insights and draft reviews. Claude AI and Gemini AI were used in the creation of this essay. I plan to post follow-ups on my Substack (https://novotia.pub).

1. B. Woltzenlogel Paleo, “Leibniz’s characteristica universalis and calculus ratiocinator today,” in Death and Anti-Death, Volume 14: Four Decades After Michael Polanyi, Three Centuries After G.W. Leibniz, Tandy, C., Ed., Ann Arbor, MI, USA: Ria Univ. Press, 2016, pp. 313–332. [Online]. Available: https://www.researchgate.net/publication/311456139

2. G. Boole, The Mathematical Analysis of Logic, Being an Essay Towards a Calculus of Deductive Reasoning. Cambridge, U.K.: Macmillan, Barclay, & Macmillan, 1847.

3. C. E. Shannon, “A symbolic analysis of relay and switching circuits,” Trans. Amer. Inst. Elect. Eng., vol. 57, no. 12, pp. 713–723, Dec. 1938, doi: 10.1109/T-AIEE.1938.5057767.

4. F. Nietzsche, “On truth and lies in a nonmoral sense,” in Truth, J. Medina and D. Wood, Eds., Hoboken, NJ, USA: Wiley, 2005, pp. 14–25.

5. M. Foucault and C. Gordon, Power/Knowledge: Selected Interviews and Other Writings, 1972-1977, 1st American ed. New York, NY, USA: Pantheon Books, 1980.

6. A. Askell, J. Carlsmith, C. Olah, J. Kaplan, and H. Karnofsky, “Claude’s constitution,” Anthropic, Jan. 22, 2026. [Online]. Available: https://www.anthropic.com/news/claude-new-constitution

7. “Gemini 3.1 Pro model card,” Google DeepMind, Feb. 2026. [Online]. Available: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-3-1-Pro-Model-Card.pdf

8. “Frontier safety framework, version 3.0,” Google DeepMind, Sep. 2025. [Online]. Available: https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/strengthening-our-frontier-safety-framework/frontier-safety-framework_3.pdf

9. “OpenAI Model Spec,” OpenAI, Dec. 18, 2025. [Online]. Available: https://model-spec.openai.com/2025-12-18.html

10. “Frontier artificial intelligence framework,” xAI, Dec. 2025. [Online]. Available: https://data.x.ai/2025-12-31-xai-frontier-artificial-intelligence-framework.pdf

11. M. Sharma et al., “Towards understanding sycophancy in language models,” 2025, arXiv:2310.13548.

12. Y. N. Harari, Sapiens [Tenth Anniversary Edition] A Brief History of Humankind. New York, NY, USA: HarperCollins, 2025.

13. D. Deutsch, The Beginning of Infinity: Explanations That Transform the World. New York, NY, USA: Penguin Books, 2014.

14. T. M. Moerland, J. Broekens, A. Plaat, and C. M. Jonker, “Model-based reinforcement learning: A survey,” 2022, arXiv:2006.16712.

15. D. Ha and J. Schmidhuber, “World models,” Zenodo, Mar. 28, 2018. doi: 10.5281/zenodo.1207631. [Online]. Available: https://zenodo.org/records/1207631

16. Y. LeCun, “A path towards autonomous machine intelligence (version 0.9.2),” OpenReview Arch., Jun. 27, 2022. [Online]. Available: https://openreview.net/forum?id=BZ5a1r-kVsf